An Entity-Centric Coreference Resolution System for Person Entities with Rich Linguistic Information

Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 741–752, Dublin, Ireland, August 23-29 2014. An Entity-Centric Coreference Resolution System for Person Entities with Rich Linguistic Information Marcos Garcia and Pablo Gamallo Centro Singular de Investigaci´on en Tecnolox´ıas da Informaci´on (CiTIUS) University of Santiago de Compostela { marcos.garcia.gonzalez, pablo.gamallo}@usc.es Abstract This paper presents a first version of LinkPeople, an entity-centric system for coreference reso- lution of person entities. The approach combines (i) a multi-pass architecture which takes advan- tage of entity features at document-level with (ii) a set of linguistically-motivated constraints and rules which allows the system to restrict the candidates of a given mention. The paper includes evaluations and error analysis of LinkPeople in 3 different languages, achieving promising results (more than 81% F1 in different metrics). Both the system and the corpora are freely distributed. 1 Introduction Coreference Resolution (CR) is a crucial task for several Natural Language Processing (NLP) applica- tions such as Text Summarization, Machine Translation or Information Extraction (IE). Specially for IE, person entities are those which more effort have deserved from different perspectives. Evaluations such as the Knowledge Base Population (KBP) Slot Filling Task (in the Text Analysis Con- ference)1 and the Person Attribute Extraction (in the Web People Search Evaluation Campaign, WePS)2, tasks such as Personal Name Matching (Cohen et al., 2003), or different works on Relation Extraction of person entities (Mann, 2002; Garcia and Gamallo, 2013) are some examples of their importance. Recently, entity-centric models for coreference resolution, which use features from all the mentions of an entity, have shown better performance than pair-mention systems, which carry out coreference resolution on single pairs of mentions (Lee et al., 2013).3 Furthermore, the use of linguistic information such as syntax or semantic knowledge has proved to be essential for high-precision CR (Ng and Cardie, 2002; Ponzetto and Strube, 2006; Uryupina, 2007). This paper presents the first version of LinkPeople, an open-source system for CR of person entities. LinkPeople is inspired by the Stanford Deterministic Coreference Resolution System (Raghunathan et al., 2010; Lee et al., 2013), using a multi-pass architecture which applies a battery of modules sorted from high-precision to high-recall. Moreover, the system presented in this paper adds new sieves based on linguistic knowledge, for both cataphoric and anaphoric mentions: It includes a high-precision module which finds cataphoric mentions of Noun Phrases (NP) and personal and elliptical pronouns. The inclusion of this module is based on the claim that definite NPs are not primarily anaphoric (Vieira and Poesio, 2000). In addition, LinkPeople applies a set of syntactic constraints on the pronominal CR module, increasing its precision by blocking links which do not satisfy the constraints (Mitkov, 1998; Palomar et al., 2001; Chaves and Rino, 2007). The system was evaluated in three languages (Portuguese, Spanish and Galician) with promising re- sults (F1 ≈ 83%, with BLANC score). Both LinkPeople and the corpora are freely distributed.4 This work is licenced under a Creative Commons Attribution 4.0 International License. Page numbers and …

Content truncated to 500 words